“e86237890ee9daa8f6ab64ed225a027ae63ebdfc”上不存在“git@gitcode.net:taosdata/tdengine.git”
operators.py 71.8 KB
Newer Older
W
wangxinxin08 已提交
1
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
Q
qingqing01 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# function:
#    operators to process sample,
#    eg: decode/resize/crop image

from __future__ import absolute_import
from __future__ import print_function
from __future__ import division

try:
    from collections.abc import Sequence
except Exception:
    from collections import Sequence

W
wangxinxin08 已提交
28
from numbers import Number, Integral
Q
qingqing01 已提交
29 30 31 32 33 34

import uuid
import random
import math
import numpy as np
import os
W
wangxinxin08 已提交
35
import copy
Q
qingqing01 已提交
36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51

import cv2
from PIL import Image, ImageEnhance, ImageDraw

from ppdet.core.workspace import serializable
from ppdet.modeling.layers import AnchorGrid

from .op_helper import (satisfy_sample_constraint, filter_and_process,
                        generate_sample_bbox, clip_bbox, data_anchor_sampling,
                        satisfy_sample_constraint_coverage, crop_image_sampling,
                        generate_sample_bbox_square, bbox_area_sampling,
                        is_poly, gaussian_radius, draw_gaussian)

from ppdet.utils.logger import setup_logger
logger = setup_logger(__name__)

W
wangxinxin08 已提交
52
registered_ops = []
Q
qingqing01 已提交
53

W
wangxinxin08 已提交
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79

def register_op(cls):
    registered_ops.append(cls.__name__)
    if not hasattr(BaseOperator, cls.__name__):
        setattr(BaseOperator, cls.__name__, cls)
    else:
        raise KeyError("The {} class has been registered.".format(cls.__name__))
    return serializable(cls)


class BboxError(ValueError):
    pass


class ImageError(ValueError):
    pass


class BaseOperator(object):
    def __init__(self, name=None):
        if name is None:
            name = self.__class__.__name__
        self._id = name + '_' + str(uuid.uuid4())[-6:]

    def apply(self, sample, context=None):
        """ Process a sample.
Q
qingqing01 已提交
80
        Args:
W
wangxinxin08 已提交
81 82 83 84
            sample (dict): a dict of sample, eg: {'image':xx, 'label': xxx}
            context (dict): info about this sample processing
        Returns:
            result (dict): a processed sample
Q
qingqing01 已提交
85
        """
W
wangxinxin08 已提交
86
        return sample
Q
qingqing01 已提交
87

W
wangxinxin08 已提交
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112
    def __call__(self, sample, context=None):
        """ Process a sample.
        Args:
            sample (dict): a dict of sample, eg: {'image':xx, 'label': xxx}
            context (dict): info about this sample processing
        Returns:
            result (dict): a processed sample
        """
        if isinstance(sample, Sequence):
            for i in range(len(sample)):
                sample[i] = self.apply(sample[i], context)
        else:
            sample = self.apply(sample, context)
        return sample

    def __str__(self):
        return str(self._id)


@register_op
class Decode(BaseOperator):
    def __init__(self):
        """ Transform the image data to numpy format following the rgb format
        """
        super(Decode, self).__init__()
Q
qingqing01 已提交
113

W
wangxinxin08 已提交
114
    def apply(self, sample, context=None):
Q
qingqing01 已提交
115 116 117 118
        """ load image if 'im_file' field is not empty but 'image' is"""
        if 'image' not in sample:
            with open(sample['im_file'], 'rb') as f:
                sample['image'] = f.read()
W
wangxinxin08 已提交
119
            sample.pop('im_file')
Q
qingqing01 已提交
120 121 122 123 124

        im = sample['image']
        data = np.frombuffer(im, dtype='uint8')
        im = cv2.imdecode(data, 1)  # BGR mode, but need RGB mode

W
wangxinxin08 已提交
125 126
        im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)

Q
qingqing01 已提交
127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144
        sample['image'] = im
        if 'h' not in sample:
            sample['h'] = im.shape[0]
        elif sample['h'] != im.shape[0]:
            logger.warn(
                "The actual image height: {} is not equal to the "
                "height: {} in annotation, and update sample['h'] by actual "
                "image height.".format(im.shape[0], sample['h']))
            sample['h'] = im.shape[0]
        if 'w' not in sample:
            sample['w'] = im.shape[1]
        elif sample['w'] != im.shape[1]:
            logger.warn(
                "The actual image width: {} is not equal to the "
                "width: {} in annotation, and update sample['w'] by actual "
                "image width.".format(im.shape[1], sample['w']))
            sample['w'] = im.shape[1]

W
wangxinxin08 已提交
145 146
        sample['im_shape'] = np.array(im.shape[:2], dtype=np.float32)
        sample['scale_factor'] = np.array([1., 1.], dtype=np.float32)
Q
qingqing01 已提交
147 148 149 150
        return sample


@register_op
W
wangxinxin08 已提交
151 152
class Permute(BaseOperator):
    def __init__(self):
Q
qingqing01 已提交
153
        """
W
wangxinxin08 已提交
154
        Change the channel to be (C, H, W)
Q
qingqing01 已提交
155
        """
W
wangxinxin08 已提交
156
        super(Permute, self).__init__()
Q
qingqing01 已提交
157

W
wangxinxin08 已提交
158
    def apply(self, sample, context=None):
Q
qingqing01 已提交
159
        im = sample['image']
W
wangxinxin08 已提交
160 161
        im = im.transpose((2, 0, 1))
        sample['image'] = im
Q
qingqing01 已提交
162 163 164 165
        return sample


@register_op
W
wangxinxin08 已提交
166 167 168 169 170 171 172 173
class Lighting(BaseOperator):
    """
    Lighting the imagen by eigenvalues and eigenvectors
    Args:
        eigval (list): eigenvalues
        eigvec (list): eigenvectors
        alphastd (float): random weight of lighting, 0.1 by default
    """
Q
qingqing01 已提交
174

W
wangxinxin08 已提交
175 176 177 178 179
    def __init__(self, eigval, eigvec, alphastd=0.1):
        super(Lighting, self).__init__()
        self.alphastd = alphastd
        self.eigval = np.array(eigval).astype('float32')
        self.eigvec = np.array(eigvec).astype('float32')
Q
qingqing01 已提交
180

W
wangxinxin08 已提交
181 182 183
    def apply(self, sample, context=None):
        alpha = np.random.normal(scale=self.alphastd, size=(3, ))
        sample['image'] += np.dot(self.eigvec, self.eigval * alpha)
Q
qingqing01 已提交
184 185 186 187
        return sample


@register_op
W
wangxinxin08 已提交
188 189
class RandomErasingImage(BaseOperator):
    def __init__(self, prob=0.5, lower=0.02, higher=0.4, aspect_ratio=0.3):
Q
qingqing01 已提交
190
        """
W
wangxinxin08 已提交
191
        Random Erasing Data Augmentation, see https://arxiv.org/abs/1708.04896
Q
qingqing01 已提交
192
        Args:
W
wangxinxin08 已提交
193 194 195 196
            prob (float): probability to carry out random erasing
            lower (float): lower limit of the erasing area ratio
            heigher (float): upper limit of the erasing area ratio
            aspect_ratio (float): aspect ratio of the erasing region
Q
qingqing01 已提交
197
        """
W
wangxinxin08 已提交
198
        super(RandomErasingImage, self).__init__()
Q
qingqing01 已提交
199
        self.prob = prob
W
wangxinxin08 已提交
200 201 202
        self.lower = lower
        self.heigher = heigher
        self.aspect_ratio = aspect_ratio
Q
qingqing01 已提交
203

W
wangxinxin08 已提交
204 205 206 207 208 209 210
    def apply(self, sample):
        gt_bbox = sample['gt_bbox']
        im = sample['image']
        if not isinstance(im, np.ndarray):
            raise TypeError("{}: image is not a numpy array.".format(self))
        if len(im.shape) != 3:
            raise ImageError("{}: image is not 3-dimensional.".format(self))
Q
qingqing01 已提交
211

W
wangxinxin08 已提交
212 213 214
        for idx in range(gt_bbox.shape[0]):
            if self.prob <= np.random.rand():
                continue
Q
qingqing01 已提交
215

W
wangxinxin08 已提交
216 217 218 219
            x1, y1, x2, y2 = gt_bbox[idx, :]
            w_bbox = x2 - x1 + 1
            h_bbox = y2 - y1 + 1
            area = w_bbox * h_bbox
Q
qingqing01 已提交
220

W
wangxinxin08 已提交
221 222 223
            target_area = random.uniform(self.lower, self.higher) * area
            aspect_ratio = random.uniform(self.aspect_ratio,
                                          1 / self.aspect_ratio)
Q
qingqing01 已提交
224

W
wangxinxin08 已提交
225 226
            h = int(round(math.sqrt(target_area * aspect_ratio)))
            w = int(round(math.sqrt(target_area / aspect_ratio)))
Q
qingqing01 已提交
227

W
wangxinxin08 已提交
228 229 230 231 232 233
            if w < w_bbox and h < h_bbox:
                off_y1 = random.randint(0, int(h_bbox - h))
                off_x1 = random.randint(0, int(w_bbox - w))
                im[int(y1 + off_y1):int(y1 + off_y1 + h), int(x1 + off_x1):int(
                    x1 + off_x1 + w), :] = 0
        sample['image'] = im
Q
qingqing01 已提交
234 235 236 237
        return sample


@register_op
W
wangxinxin08 已提交
238 239 240
class NormalizeImage(BaseOperator):
    def __init__(self, mean=[0.485, 0.456, 0.406], std=[1, 1, 1],
                 is_scale=True):
Q
qingqing01 已提交
241 242
        """
        Args:
W
wangxinxin08 已提交
243 244
            mean (list): the pixel mean
            std (list): the pixel variance
Q
qingqing01 已提交
245
        """
W
wangxinxin08 已提交
246 247 248 249 250 251 252 253 254 255
        super(NormalizeImage, self).__init__()
        self.mean = mean
        self.std = std
        self.is_scale = is_scale
        if not (isinstance(self.mean, list) and isinstance(self.std, list) and
                isinstance(self.is_scale, bool)):
            raise TypeError("{}: input type is invalid.".format(self))
        from functools import reduce
        if reduce(lambda x, y: x * y, self.std) == 0:
            raise ValueError('{}: std is invalid!'.format(self))
Q
qingqing01 已提交
256

W
wangxinxin08 已提交
257 258 259 260 261 262 263 264 265 266
    def apply(self, sample, context=None):
        """Normalize the image.
        Operators:
            1.(optional) Scale the image to [0,1]
            2. Each pixel minus mean and is divided by std
        """
        im = sample['image']
        im = im.astype(np.float32, copy=False)
        mean = np.array(self.mean)[np.newaxis, np.newaxis, :]
        std = np.array(self.std)[np.newaxis, np.newaxis, :]
Q
qingqing01 已提交
267

W
wangxinxin08 已提交
268 269
        if self.is_scale:
            im = im / 255.0
Q
qingqing01 已提交
270

W
wangxinxin08 已提交
271 272
        im -= mean
        im /= std
Q
qingqing01 已提交
273

W
wangxinxin08 已提交
274
        sample['image'] = im
Q
qingqing01 已提交
275 276 277 278
        return sample


@register_op
W
wangxinxin08 已提交
279
class GridMask(BaseOperator):
Q
qingqing01 已提交
280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300
    def __init__(self,
                 use_h=True,
                 use_w=True,
                 rotate=1,
                 offset=False,
                 ratio=0.5,
                 mode=1,
                 prob=0.7,
                 upper_iter=360000):
        """
        GridMask Data Augmentation, see https://arxiv.org/abs/2001.04086
        Args:
            use_h (bool): whether to mask vertically
            use_w (boo;): whether to mask horizontally
            rotate (float): angle for the mask to rotate
            offset (float): mask offset
            ratio (float): mask ratio
            mode (int): gridmask mode
            prob (float): max probability to carry out gridmask
            upper_iter (int): suggested to be equal to global max_iter
        """
W
wangxinxin08 已提交
301
        super(GridMask, self).__init__()
Q
qingqing01 已提交
302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321
        self.use_h = use_h
        self.use_w = use_w
        self.rotate = rotate
        self.offset = offset
        self.ratio = ratio
        self.mode = mode
        self.prob = prob
        self.upper_iter = upper_iter

        from .gridmask_utils import GridMask
        self.gridmask_op = GridMask(
            use_h,
            use_w,
            rotate=rotate,
            offset=offset,
            ratio=ratio,
            mode=mode,
            prob=prob,
            upper_iter=upper_iter)

W
wangxinxin08 已提交
322 323
    def apply(self, sample, context=None):
        sample['image'] = self.gridmask_op(sample['image'], sample['curr_iter'])
Q
qingqing01 已提交
324 325 326 327
        return sample


@register_op
W
wangxinxin08 已提交
328 329 330 331 332 333 334 335 336 337 338 339
class RandomDistort(BaseOperator):
    """Random color distortion.
    Args:
        hue (list): hue settings. in [lower, upper, probability] format.
        saturation (list): saturation settings. in [lower, upper, probability] format.
        contrast (list): contrast settings. in [lower, upper, probability] format.
        brightness (list): brightness settings. in [lower, upper, probability] format.
        random_apply (bool): whether to apply in random (yolo) or fixed (SSD)
            order.
        count (int): the number of doing distrot
        random_channel (bool): whether to swap channels randomly
    """
Q
qingqing01 已提交
340

W
wangxinxin08 已提交
341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
    def __init__(self,
                 hue=[-18, 18, 0.5],
                 saturation=[0.5, 1.5, 0.5],
                 contrast=[0.5, 1.5, 0.5],
                 brightness=[0.5, 1.5, 0.5],
                 random_apply=True,
                 count=4,
                 random_channel=False):
        super(RandomDistort, self).__init__()
        self.hue = hue
        self.saturation = saturation
        self.contrast = contrast
        self.brightness = brightness
        self.random_apply = random_apply
        self.count = count
        self.random_channel = random_channel
Q
qingqing01 已提交
357

W
wangxinxin08 已提交
358 359 360 361
    def apply_hue(self, img):
        low, high, prob = self.hue
        if np.random.uniform(0., 1.) < prob:
            return img
Q
qingqing01 已提交
362

W
wangxinxin08 已提交
363 364 365 366 367 368 369 370 371 372 373 374
        img = img.astype(np.float32)
        # it works, but result differ from HSV version
        delta = np.random.uniform(low, high)
        u = np.cos(delta * np.pi)
        w = np.sin(delta * np.pi)
        bt = np.array([[1.0, 0.0, 0.0], [0.0, u, -w], [0.0, w, u]])
        tyiq = np.array([[0.299, 0.587, 0.114], [0.596, -0.274, -0.321],
                         [0.211, -0.523, 0.311]])
        ityiq = np.array([[1.0, 0.956, 0.621], [1.0, -0.272, -0.647],
                          [1.0, -1.107, 1.705]])
        t = np.dot(np.dot(ityiq, bt), tyiq).T
        img = np.dot(img, t)
Q
qingqing01 已提交
375 376
        return img

W
wangxinxin08 已提交
377 378 379 380 381 382 383 384 385 386 387 388
    def apply_saturation(self, img):
        low, high, prob = self.saturation
        if np.random.uniform(0., 1.) < prob:
            return img
        delta = np.random.uniform(low, high)
        img = img.astype(np.float32)
        # it works, but result differ from HSV version
        gray = img * np.array([[[0.299, 0.587, 0.114]]], dtype=np.float32)
        gray = gray.sum(axis=2, keepdims=True)
        gray *= (1.0 - delta)
        img *= delta
        img += gray
Q
qingqing01 已提交
389 390
        return img

W
wangxinxin08 已提交
391 392 393 394 395 396 397
    def apply_contrast(self, img):
        low, high, prob = self.contrast
        if np.random.uniform(0., 1.) < prob:
            return img
        delta = np.random.uniform(low, high)
        img = img.astype(np.float32)
        img *= delta
Q
qingqing01 已提交
398 399
        return img

W
wangxinxin08 已提交
400 401 402 403 404 405 406 407
    def apply_brightness(self, img):
        low, high, prob = self.brightness
        if np.random.uniform(0., 1.) < prob:
            return img
        delta = np.random.uniform(low, high)
        img = img.astype(np.float32)
        img += delta
        return img
Q
qingqing01 已提交
408

W
wangxinxin08 已提交
409 410 411 412 413 414 415 416 417 418 419 420
    def apply(self, sample, context=None):
        img = sample['image']
        if self.random_apply:
            functions = [
                self.apply_brightness, self.apply_contrast,
                self.apply_saturation, self.apply_hue
            ]
            distortions = np.random.permutation(functions)[:self.count]
            for func in distortions:
                img = func(img)
            sample['image'] = img
            return sample
Q
qingqing01 已提交
421

W
wangxinxin08 已提交
422 423
        img = self.apply_brightness(img)
        mode = np.random.randint(0, 2)
Q
qingqing01 已提交
424

W
wangxinxin08 已提交
425 426
        if mode:
            img = self.apply_contrast(img)
Q
qingqing01 已提交
427

W
wangxinxin08 已提交
428 429
        img = self.apply_saturation(img)
        img = self.apply_hue(img)
Q
qingqing01 已提交
430

W
wangxinxin08 已提交
431 432
        if not mode:
            img = self.apply_contrast(img)
Q
qingqing01 已提交
433

W
wangxinxin08 已提交
434 435 436 437
        if self.random_channel:
            if np.random.randint(0, 2):
                img = img[..., np.random.permutation(3)]
        sample['image'] = img
Q
qingqing01 已提交
438 439 440 441
        return sample


@register_op
W
wangxinxin08 已提交
442 443
class AutoAugment(BaseOperator):
    def __init__(self, autoaug_type="v1"):
Q
qingqing01 已提交
444 445
        """
        Args:
W
wangxinxin08 已提交
446
            autoaug_type (str): autoaug type, support v0, v1, v2, v3, test
Q
qingqing01 已提交
447
        """
W
wangxinxin08 已提交
448 449
        super(AutoAugment, self).__init__()
        self.autoaug_type = autoaug_type
Q
qingqing01 已提交
450

W
wangxinxin08 已提交
451
    def apply(self, sample, context=None):
Q
qingqing01 已提交
452
        """
W
wangxinxin08 已提交
453 454 455 456 457 458 459 460 461
        Learning Data Augmentation Strategies for Object Detection, see https://arxiv.org/abs/1906.11172
        """
        im = sample['image']
        gt_bbox = sample['gt_bbox']
        if not isinstance(im, np.ndarray):
            raise TypeError("{}: image is not a numpy array.".format(self))
        if len(im.shape) != 3:
            raise ImageError("{}: image is not 3-dimensional.".format(self))
        if len(gt_bbox) == 0:
Q
qingqing01 已提交
462 463
            return sample

W
wangxinxin08 已提交
464 465 466 467 468 469
        height, width, _ = im.shape
        norm_gt_bbox = np.ones_like(gt_bbox, dtype=np.float32)
        norm_gt_bbox[:, 0] = gt_bbox[:, 1] / float(height)
        norm_gt_bbox[:, 1] = gt_bbox[:, 0] / float(width)
        norm_gt_bbox[:, 2] = gt_bbox[:, 3] / float(height)
        norm_gt_bbox[:, 3] = gt_bbox[:, 2] / float(width)
Q
qingqing01 已提交
470

W
wangxinxin08 已提交
471 472 473 474 475 476 477 478
        from .autoaugment_utils import distort_image_with_autoaugment
        im, norm_gt_bbox = distort_image_with_autoaugment(im, norm_gt_bbox,
                                                          self.autoaug_type)

        gt_bbox[:, 0] = norm_gt_bbox[:, 1] * float(width)
        gt_bbox[:, 1] = norm_gt_bbox[:, 0] * float(height)
        gt_bbox[:, 2] = norm_gt_bbox[:, 3] * float(width)
        gt_bbox[:, 3] = norm_gt_bbox[:, 2] * float(height)
Q
qingqing01 已提交
479 480 481 482 483 484 485

        sample['image'] = im
        sample['gt_bbox'] = gt_bbox
        return sample


@register_op
W
wangxinxin08 已提交
486 487
class RandomFlip(BaseOperator):
    def __init__(self, prob=0.5):
Q
qingqing01 已提交
488 489
        """
        Args:
W
wangxinxin08 已提交
490
            prob (float): the probability of flipping image
Q
qingqing01 已提交
491
        """
W
wangxinxin08 已提交
492 493 494 495
        super(RandomFlip, self).__init__()
        self.prob = prob
        if not (isinstance(self.prob, float)):
            raise TypeError("{}: input type is invalid.".format(self))
Q
qingqing01 已提交
496

W
wangxinxin08 已提交
497 498 499 500 501
    def apply_segm(self, segms, height, width):
        def _flip_poly(poly, width):
            flipped_poly = np.array(poly)
            flipped_poly[0::2] = width - np.array(poly[0::2])
            return flipped_poly.tolist()
Q
qingqing01 已提交
502

W
wangxinxin08 已提交
503 504 505 506 507 508 509
        def _flip_rle(rle, height, width):
            if 'counts' in rle and type(rle['counts']) == list:
                rle = mask_util.frPyObjects(rle, height, width)
            mask = mask_util.decode(rle)
            mask = mask[:, ::-1]
            rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
            return rle
Q
qingqing01 已提交
510

W
wangxinxin08 已提交
511 512 513 514 515 516 517 518 519 520
        flipped_segms = []
        for segm in segms:
            if is_poly(segm):
                # Polygon format
                flipped_segms.append([_flip_poly(poly, width) for poly in segm])
            else:
                # RLE format
                import pycocotools.mask as mask_util
                flipped_segms.append(_flip_rle(segm, height, width))
        return flipped_segms
Q
qingqing01 已提交
521

W
wangxinxin08 已提交
522 523 524 525 526 527
    def apply_keypoint(self, gt_keypoint, width):
        for i in range(gt_keypoint.shape[1]):
            if i % 2 == 0:
                old_x = gt_keypoint[:, i].copy()
                gt_keypoint[:, i] = width - old_x
        return gt_keypoint
Q
qingqing01 已提交
528

W
wangxinxin08 已提交
529 530
    def apply_image(self, image):
        return image[:, ::-1, :]
Q
qingqing01 已提交
531

W
wangxinxin08 已提交
532 533 534 535 536 537
    def apply_bbox(self, bbox, width):
        oldx1 = bbox[:, 0].copy()
        oldx2 = bbox[:, 2].copy()
        bbox[:, 0] = width - oldx2
        bbox[:, 2] = width - oldx1
        return bbox
Q
qingqing01 已提交
538

W
wangxinxin08 已提交
539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571
    def apply(self, sample, context=None):
        """Filp the image and bounding box.
        Operators:
            1. Flip the image numpy.
            2. Transform the bboxes' x coordinates.
              (Must judge whether the coordinates are normalized!)
            3. Transform the segmentations' x coordinates.
              (Must judge whether the coordinates are normalized!)
        Output:
            sample: the image, bounding box and segmentation part
                    in sample are flipped.
        """
        if np.random.uniform(0, 1) < self.prob:
            im = sample['image']
            height, width = im.shape[:2]
            im = self.apply_image(im)
            if 'gt_bbox' in sample and len(sample['gt_bbox']) > 0:
                sample['gt_bbox'] = self.apply_bbox(sample['gt_bbox'], width)
            if 'gt_poly' in sample and len(sample['gt_poly']) > 0:
                sample['gt_poly'] = self.apply_segm(sample['gt_poly'], height,
                                                    width)
            if 'gt_keypoint' in sample and len(sample['gt_keypoint']) > 0:
                sample['gt_keypoint'] = self.apply_keypoint(
                    sample['gt_keypoint'], width)

            if 'semantic' in sample and sample['semantic']:
                sample['semantic'] = sample['semantic'][:, ::-1]

            if 'gt_segm' in sample and sample['gt_segm'].any():
                sample['gt_segm'] = sample['gt_segm'][:, :, ::-1]

            sample['flipped'] = True
            sample['image'] = im
Q
qingqing01 已提交
572 573 574 575
        return sample


@register_op
W
wangxinxin08 已提交
576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596
class Resize(BaseOperator):
    def __init__(self, target_size, keep_ratio, interp=cv2.INTER_LINEAR):
        """
        Resize image to target size. if keep_ratio is True, 
        resize the image's long side to the maximum of target_size
        if keep_ratio is False, resize the image to target size(h, w)
        Args:
            target_size (int|list): image target size
            keep_ratio (bool): whether keep_ratio or not, default true
            interp (int): the interpolation method
        """
        super(Resize, self).__init__()
        self.keep_ratio = keep_ratio
        self.interp = interp
        if not isinstance(target_size, (Integral, Sequence)):
            raise TypeError(
                "Type of target_size is invalid. Must be Integer or List or Tuple, now is {}".
                format(type(target_size)))
        if isinstance(target_size, Integral):
            target_size = [target_size, target_size]
        self.target_size = target_size
Q
qingqing01 已提交
597

W
wangxinxin08 已提交
598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627
    def apply_image(self, image, scale):
        im_scale_x, im_scale_y = scale

        return cv2.resize(
            image,
            None,
            None,
            fx=im_scale_x,
            fy=im_scale_y,
            interpolation=self.interp)

    def apply_bbox(self, bbox, scale, size):
        im_scale_x, im_scale_y = scale
        resize_w, resize_h = size
        bbox[:, 0::2] *= im_scale_x
        bbox[:, 1::2] *= im_scale_y
        bbox[:, 0::2] = np.clip(bbox[:, 0::2], 0, resize_w)
        bbox[:, 1::2] = np.clip(bbox[:, 1::2], 0, resize_h)
        return bbox

    def apply_segm(self, segms, im_size, scale):
        def _resize_poly(poly, im_scale_x, im_scale_y):
            resized_poly = np.array(poly)
            resized_poly[0::2] *= im_scale_x
            resized_poly[1::2] *= im_scale_y
            return resized_poly.tolist()

        def _resize_rle(rle, im_h, im_w, im_scale_x, im_scale_y):
            if 'counts' in rle and type(rle['counts']) == list:
                rle = mask_util.frPyObjects(rle, im_h, im_w)
Q
qingqing01 已提交
628

W
wangxinxin08 已提交
629 630 631 632 633 634 635 636 637 638
            mask = mask_util.decode(rle)
            mask = cv2.resize(
                image,
                None,
                None,
                fx=im_scale_x,
                fy=im_scale_y,
                interpolation=self.interp)
            rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
            return rle
Q
qingqing01 已提交
639

W
wangxinxin08 已提交
640 641 642 643 644 645 646 647 648 649 650 651 652 653
        im_h, im_w = im_size
        im_scale_x, im_scale_y = scale
        resized_segms = []
        for segm in segms:
            if is_poly(segm):
                # Polygon format
                resized_segms.append([
                    _resize_poly(poly, im_scale_x, im_scale_y) for poly in segm
                ])
            else:
                # RLE format
                import pycocotools.mask as mask_util
                resized_segms.append(
                    _resize_rle(segm, im_h, im_w, im_scale_x, im_scale_y))
Q
qingqing01 已提交
654

W
wangxinxin08 已提交
655
        return resized_segms
Q
qingqing01 已提交
656

W
wangxinxin08 已提交
657 658 659 660 661 662 663 664
    def apply(self, sample, context=None):
        """ Resize the image numpy.
        """
        im = sample['image']
        if not isinstance(im, np.ndarray):
            raise TypeError("{}: image type is not numpy.".format(self))
        if len(im.shape) != 3:
            raise ImageError('{}: image is not 3-dimensional.'.format(self))
Q
qingqing01 已提交
665

W
wangxinxin08 已提交
666 667 668
        # apply image
        im_shape = im.shape
        if self.keep_ratio:
Q
qingqing01 已提交
669

W
wangxinxin08 已提交
670 671
            im_size_min = np.min(im_shape[0:2])
            im_size_max = np.max(im_shape[0:2])
Q
qingqing01 已提交
672

W
wangxinxin08 已提交
673 674
            target_size_min = np.min(self.target_size)
            target_size_max = np.max(self.target_size)
Q
qingqing01 已提交
675

W
wangxinxin08 已提交
676 677
            im_scale = min(target_size_min / im_size_min,
                           target_size_max / im_size_max)
Q
qingqing01 已提交
678

W
wangxinxin08 已提交
679 680
            resize_h = im_scale * float(im_shape[0])
            resize_w = im_scale * float(im_shape[1])
Q
qingqing01 已提交
681

W
wangxinxin08 已提交
682 683
            im_scale_x = im_scale
            im_scale_y = im_scale
Q
qingqing01 已提交
684
        else:
W
wangxinxin08 已提交
685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738
            resize_h, resize_w = self.target_size
            im_scale_y = resize_h / im_shape[0]
            im_scale_x = resize_w / im_shape[1]

        im = self.apply_image(sample['image'], [im_scale_x, im_scale_y])
        sample['image'] = im
        sample['im_shape'] = np.asarray([resize_h, resize_w], dtype=np.float32)
        if 'scale_factor' in sample:
            scale_factor = sample['scale_factor']
            sample['scale_factor'] = np.asarray(
                [scale_factor[0] * im_scale_y, scale_factor[1] * im_scale_x],
                dtype=np.float32)
        else:
            sample['scale_factor'] = np.asarray(
                [im_scale_y, im_scale_x], dtype=np.float32)

        # apply bbox
        if 'gt_bbox' in sample and len(sample['gt_bbox']) > 0:
            sample['gt_bbox'] = self.apply_bbox(sample['gt_bbox'],
                                                [im_scale_x, im_scale_y],
                                                [resize_w, resize_h])

        # apply polygon
        if 'gt_poly' in sample and len(sample['gt_poly']) > 0:
            sample['gt_poly'] = self.apply_segm(sample['gt_poly'], im_shape[:2],
                                                [im_scale_x, im_scale_y])

        # apply semantic
        if 'semantic' in sample and sample['semantic']:
            semantic = sample['semantic']
            semantic = cv2.resize(
                semantic.astype('float32'),
                None,
                None,
                fx=im_scale_x,
                fy=im_scale_y,
                interpolation=self.interp)
            semantic = np.asarray(semantic).astype('int32')
            semantic = np.expand_dims(semantic, 0)
            sample['semantic'] = semantic

        # apply gt_segm
        if 'gt_segm' in sample and len(sample['gt_segm']) > 0:
            masks = [
                cv2.resize(
                    gt_segm,
                    None,
                    None,
                    fx=im_scale_x,
                    fy=im_scale_y,
                    interpolation=cv2.INTER_NEAREST)
                for gt_segm in sample['gt_segm']
            ]
            sample['gt_segm'] = np.asarray(masks).astype(np.uint8)
Q
qingqing01 已提交
739 740 741 742 743

        return sample


@register_op
W
wangxinxin08 已提交
744
class MultiscaleTestResize(BaseOperator):
Q
qingqing01 已提交
745
    def __init__(self,
W
wangxinxin08 已提交
746 747 748 749 750 751 752 753 754 755 756 757 758 759 760
                 origin_target_size=[800, 1333],
                 target_size=[],
                 interp=cv2.INTER_LINEAR,
                 use_flip=True):
        """
        Rescale image to the each size in target size, and capped at max_size.
        Args:
            origin_target_size (list): origin target size of image
            target_size (list): A list of target sizes of image.
            interp (int): the interpolation method.
            use_flip (bool): whether use flip augmentation.
        """
        super(MultiscaleTestResize, self).__init__()
        self.interp = interp
        self.use_flip = use_flip
Q
qingqing01 已提交
761

W
wangxinxin08 已提交
762 763 764 765 766
        if not isinstance(target_size, Sequence):
            raise TypeError(
                "Type of target_size is invalid. Must be List or Tuple, now is {}".
                format(type(target_size)))
        self.target_size = target_size
Q
qingqing01 已提交
767

W
wangxinxin08 已提交
768 769 770 771
        if not isinstance(origin_target_size, Sequence):
            raise TypeError(
                "Type of origin_target_size is invalid. Must be List or Tuple, now is {}".
                format(type(origin_target_size)))
Q
qingqing01 已提交
772

W
wangxinxin08 已提交
773
        self.origin_target_size = origin_target_size
Q
qingqing01 已提交
774

W
wangxinxin08 已提交
775 776 777 778 779 780 781 782 783 784
    def apply(self, sample, context=None):
        """ Resize the image numpy for multi-scale test.
        """
        samples = []
        resizer = Resize(
            self.origin_target_size, keep_ratio=True, interp=self.interp)
        samples.append(resizer(sample.copy(), context))
        if self.use_flip:
            flipper = RandomFlip(1.1)
            samples.append(flipper(sample.copy(), context=context))
Q
qingqing01 已提交
785

W
wangxinxin08 已提交
786 787 788
        for size in self.target_size:
            resizer = Resize(size, keep_ratio=True, interp=self.interp)
            samples.append(resizer(sample.copy(), context))
Q
qingqing01 已提交
789

W
wangxinxin08 已提交
790
        return samples
Q
qingqing01 已提交
791 792


W
wangxinxin08 已提交
793 794
@register_op
class RandomResize(BaseOperator):
Q
qingqing01 已提交
795
    def __init__(self,
W
wangxinxin08 已提交
796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828
                 target_size,
                 keep_ratio=True,
                 interp=cv2.INTER_LINEAR,
                 random_size=True,
                 random_interp=False):
        """
        Resize image to target size randomly. random target_size and interpolation method
        Args:
            target_size (int, list, tuple): image target size, if random size is True, must be list or tuple
            keep_ratio (bool): whether keep_raio or not, default true
            interp (int): the interpolation method
            random_size (bool): whether random select target size of image
            random_interp (bool): whether random select interpolation method
        """
        super(RandomResize, self).__init__()
        self.keep_ratio = keep_ratio
        self.interp = interp
        self.interps = [
            cv2.INTER_NEAREST,
            cv2.INTER_LINEAR,
            cv2.INTER_AREA,
            cv2.INTER_CUBIC,
            cv2.INTER_LANCZOS4,
        ]
        assert isinstance(target_size, (
            Integral, Sequence)), "target_size must be Integer, List or Tuple"
        if random_size and not isinstance(target_size, Sequence):
            raise TypeError(
                "Type of target_size is invalid when random_size is True. Must be List or Tuple, now is {}".
                format(type(target_size)))
        self.target_size = target_size
        self.random_size = random_size
        self.random_interp = random_interp
Q
qingqing01 已提交
829

W
wangxinxin08 已提交
830 831 832 833 834 835 836
    def apply(self, sample, context=None):
        """ Resize the image numpy.
        """
        if self.random_size:
            target_size = random.choice(self.target_size)
        else:
            target_size = self.target_size
Q
qingqing01 已提交
837

W
wangxinxin08 已提交
838 839 840 841 842 843 844
        if self.random_interp:
            interp = random.choice(self.interps)
        else:
            interp = self.interp

        resizer = Resize(target_size, self.keep_ratio, interp)
        return resizer(sample, context=context)
Q
qingqing01 已提交
845 846 847 848 849 850 851 852 853 854 855


@register_op
class RandomExpand(BaseOperator):
    """Random expand the canvas.
    Args:
        ratio (float): maximum expansion ratio.
        prob (float): probability to expand.
        fill_value (list): color value used to fill the canvas. in RGB order.
    """

W
wangxinxin08 已提交
856
    def __init__(self, ratio=4., prob=0.5, fill_value=(127.5, 127.5, 127.5)):
Q
qingqing01 已提交
857 858 859 860 861 862 863 864 865 866 867 868
        super(RandomExpand, self).__init__()
        assert ratio > 1.01, "expand ratio must be larger than 1.01"
        self.ratio = ratio
        self.prob = prob
        assert isinstance(fill_value, (Number, Sequence)), \
            "fill value must be either float or sequence"
        if isinstance(fill_value, Number):
            fill_value = (fill_value, ) * 3
        if not isinstance(fill_value, tuple):
            fill_value = tuple(fill_value)
        self.fill_value = fill_value

W
wangxinxin08 已提交
869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996
    def apply(self, sample, context=None):
        if np.random.uniform(0., 1.) < self.prob:
            return sample

        im = sample['image']
        height, width = im.shape[:2]
        ratio = np.random.uniform(1., self.ratio)
        h = int(height * ratio)
        w = int(width * ratio)
        if not h > height or not w > width:
            return sample
        y = np.random.randint(0, h - height)
        x = np.random.randint(0, w - width)
        offsets, size = [x, y], [h, w]

        pad = Pad(size,
                  pad_mode=-1,
                  offsets=offsets,
                  fill_value=self.fill_value)

        return pad(sample, context=context)


@register_op
class CropWithSampling(BaseOperator):
    def __init__(self, batch_sampler, satisfy_all=False, avoid_no_bbox=True):
        """
        Args:
            batch_sampler (list): Multiple sets of different
                                  parameters for cropping.
            satisfy_all (bool): whether all boxes must satisfy.
            e.g.[[1, 1, 1.0, 1.0, 1.0, 1.0, 0.0, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.1, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.3, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.5, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.7, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.9, 1.0],
                 [1, 50, 0.3, 1.0, 0.5, 2.0, 0.0, 1.0]]
           [max sample, max trial, min scale, max scale,
            min aspect ratio, max aspect ratio,
            min overlap, max overlap]
            avoid_no_bbox (bool): whether to to avoid the
                                  situation where the box does not appear.
        """
        super(CropWithSampling, self).__init__()
        self.batch_sampler = batch_sampler
        self.satisfy_all = satisfy_all
        self.avoid_no_bbox = avoid_no_bbox

    def apply(self, sample, context):
        """
        Crop the image and modify bounding box.
        Operators:
            1. Scale the image width and height.
            2. Crop the image according to a radom sample.
            3. Rescale the bounding box.
            4. Determine if the new bbox is satisfied in the new image.
        Returns:
            sample: the image, bounding box are replaced.
        """
        assert 'image' in sample, "image data not found"
        im = sample['image']
        gt_bbox = sample['gt_bbox']
        gt_class = sample['gt_class']
        im_height, im_width = im.shape[:2]
        gt_score = None
        if 'gt_score' in sample:
            gt_score = sample['gt_score']
        sampled_bbox = []
        gt_bbox = gt_bbox.tolist()
        for sampler in self.batch_sampler:
            found = 0
            for i in range(sampler[1]):
                if found >= sampler[0]:
                    break
                sample_bbox = generate_sample_bbox(sampler)
                if satisfy_sample_constraint(sampler, sample_bbox, gt_bbox,
                                             self.satisfy_all):
                    sampled_bbox.append(sample_bbox)
                    found = found + 1
        im = np.array(im)
        while sampled_bbox:
            idx = int(np.random.uniform(0, len(sampled_bbox)))
            sample_bbox = sampled_bbox.pop(idx)
            sample_bbox = clip_bbox(sample_bbox)
            crop_bbox, crop_class, crop_score = \
                filter_and_process(sample_bbox, gt_bbox, gt_class, scores=gt_score)
            if self.avoid_no_bbox:
                if len(crop_bbox) < 1:
                    continue
            xmin = int(sample_bbox[0] * im_width)
            xmax = int(sample_bbox[2] * im_width)
            ymin = int(sample_bbox[1] * im_height)
            ymax = int(sample_bbox[3] * im_height)
            im = im[ymin:ymax, xmin:xmax]
            sample['image'] = im
            sample['gt_bbox'] = crop_bbox
            sample['gt_class'] = crop_class
            sample['gt_score'] = crop_score
            return sample
        return sample


@register_op
class CropWithDataAchorSampling(BaseOperator):
    def __init__(self,
                 batch_sampler,
                 anchor_sampler=None,
                 target_size=None,
                 das_anchor_scales=[16, 32, 64, 128],
                 sampling_prob=0.5,
                 min_size=8.,
                 avoid_no_bbox=True):
        """
        Args:
            anchor_sampler (list): anchor_sampling sets of different
                                  parameters for cropping.
            batch_sampler (list): Multiple sets of different
                                  parameters for cropping.
              e.g.[[1, 10, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.2, 0.0]]
                  [[1, 50, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0],
                   [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0],
                   [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0],
                   [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0],
                   [1, 50, 0.3, 1.0, 1.0, 1.0, 0.0, 0.0, 1.0, 0.0]]
              [max sample, max trial, min scale, max scale,
               min aspect ratio, max aspect ratio,
               min overlap, max overlap, min coverage, max coverage]
997
            target_size (int): target image size.
W
wangxinxin08 已提交
998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028
            das_anchor_scales (list[float]): a list of anchor scales in data
                anchor smapling.
            min_size (float): minimum size of sampled bbox.
            avoid_no_bbox (bool): whether to to avoid the
                                  situation where the box does not appear.
        """
        super(CropWithDataAchorSampling, self).__init__()
        self.anchor_sampler = anchor_sampler
        self.batch_sampler = batch_sampler
        self.target_size = target_size
        self.sampling_prob = sampling_prob
        self.min_size = min_size
        self.avoid_no_bbox = avoid_no_bbox
        self.das_anchor_scales = np.array(das_anchor_scales)

    def apply(self, sample, context):
        """
        Crop the image and modify bounding box.
        Operators:
            1. Scale the image width and height.
            2. Crop the image according to a radom sample.
            3. Rescale the bounding box.
            4. Determine if the new bbox is satisfied in the new image.
        Returns:
            sample: the image, bounding box are replaced.
        """
        assert 'image' in sample, "image data not found"
        im = sample['image']
        gt_bbox = sample['gt_bbox']
        gt_class = sample['gt_class']
        image_height, image_width = im.shape[:2]
1029 1030 1031 1032
        gt_bbox[:, 0] /= image_width
        gt_bbox[:, 1] /= image_height
        gt_bbox[:, 2] /= image_width
        gt_bbox[:, 3] /= image_height
W
wangxinxin08 已提交
1033 1034 1035 1036 1037
        gt_score = None
        if 'gt_score' in sample:
            gt_score = sample['gt_score']
        sampled_bbox = []
        gt_bbox = gt_bbox.tolist()
Q
qingqing01 已提交
1038

W
wangxinxin08 已提交
1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059
        prob = np.random.uniform(0., 1.)
        if prob > self.sampling_prob:  # anchor sampling
            assert self.anchor_sampler
            for sampler in self.anchor_sampler:
                found = 0
                for i in range(sampler[1]):
                    if found >= sampler[0]:
                        break
                    sample_bbox = data_anchor_sampling(
                        gt_bbox, image_width, image_height,
                        self.das_anchor_scales, self.target_size)
                    if sample_bbox == 0:
                        break
                    if satisfy_sample_constraint_coverage(sampler, sample_bbox,
                                                          gt_bbox):
                        sampled_bbox.append(sample_bbox)
                        found = found + 1
            im = np.array(im)
            while sampled_bbox:
                idx = int(np.random.uniform(0, len(sampled_bbox)))
                sample_bbox = sampled_bbox.pop(idx)
Q
qingqing01 已提交
1060

W
wangxinxin08 已提交
1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073
                if 'gt_keypoint' in sample.keys():
                    keypoints = (sample['gt_keypoint'],
                                 sample['keypoint_ignore'])
                    crop_bbox, crop_class, crop_score, gt_keypoints = \
                        filter_and_process(sample_bbox, gt_bbox, gt_class,
                                scores=gt_score,
                                keypoints=keypoints)
                else:
                    crop_bbox, crop_class, crop_score = filter_and_process(
                        sample_bbox, gt_bbox, gt_class, scores=gt_score)
                crop_bbox, crop_class, crop_score = bbox_area_sampling(
                    crop_bbox, crop_class, crop_score, self.target_size,
                    self.min_size)
Q
qingqing01 已提交
1074

W
wangxinxin08 已提交
1075 1076 1077 1078 1079
                if self.avoid_no_bbox:
                    if len(crop_bbox) < 1:
                        continue
                im = crop_image_sampling(im, sample_bbox, image_width,
                                         image_height, self.target_size)
1080 1081 1082 1083 1084
                height, width = im.shape[:2]
                crop_bbox[:, 0] *= width
                crop_bbox[:, 1] *= height
                crop_bbox[:, 2] *= width
                crop_bbox[:, 3] *= height
W
wangxinxin08 已提交
1085 1086 1087
                sample['image'] = im
                sample['gt_bbox'] = crop_bbox
                sample['gt_class'] = crop_class
1088 1089
                if 'gt_score' in sample:
                    sample['gt_score'] = crop_score
W
wangxinxin08 已提交
1090 1091 1092 1093
                if 'gt_keypoint' in sample.keys():
                    sample['gt_keypoint'] = gt_keypoints[0]
                    sample['keypoint_ignore'] = gt_keypoints[1]
                return sample
Q
qingqing01 已提交
1094 1095
            return sample

W
wangxinxin08 已提交
1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112
        else:
            for sampler in self.batch_sampler:
                found = 0
                for i in range(sampler[1]):
                    if found >= sampler[0]:
                        break
                    sample_bbox = generate_sample_bbox_square(
                        sampler, image_width, image_height)
                    if satisfy_sample_constraint_coverage(sampler, sample_bbox,
                                                          gt_bbox):
                        sampled_bbox.append(sample_bbox)
                        found = found + 1
            im = np.array(im)
            while sampled_bbox:
                idx = int(np.random.uniform(0, len(sampled_bbox)))
                sample_bbox = sampled_bbox.pop(idx)
                sample_bbox = clip_bbox(sample_bbox)
Q
qingqing01 已提交
1113

W
wangxinxin08 已提交
1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127
                if 'gt_keypoint' in sample.keys():
                    keypoints = (sample['gt_keypoint'],
                                 sample['keypoint_ignore'])
                    crop_bbox, crop_class, crop_score, gt_keypoints = \
                        filter_and_process(sample_bbox, gt_bbox, gt_class,
                                scores=gt_score,
                                keypoints=keypoints)
                else:
                    crop_bbox, crop_class, crop_score = filter_and_process(
                        sample_bbox, gt_bbox, gt_class, scores=gt_score)
                # sampling bbox according the bbox area
                crop_bbox, crop_class, crop_score = bbox_area_sampling(
                    crop_bbox, crop_class, crop_score, self.target_size,
                    self.min_size)
Q
qingqing01 已提交
1128

W
wangxinxin08 已提交
1129 1130 1131 1132 1133 1134 1135 1136
                if self.avoid_no_bbox:
                    if len(crop_bbox) < 1:
                        continue
                xmin = int(sample_bbox[0] * image_width)
                xmax = int(sample_bbox[2] * image_width)
                ymin = int(sample_bbox[1] * image_height)
                ymax = int(sample_bbox[3] * image_height)
                im = im[ymin:ymax, xmin:xmax]
1137 1138 1139 1140 1141
                height, width = im.shape[:2]
                crop_bbox[:, 0] *= width
                crop_bbox[:, 1] *= height
                crop_bbox[:, 2] *= width
                crop_bbox[:, 3] *= height
W
wangxinxin08 已提交
1142 1143 1144
                sample['image'] = im
                sample['gt_bbox'] = crop_bbox
                sample['gt_class'] = crop_class
1145 1146
                if 'gt_score' in sample:
                    sample['gt_score'] = crop_score
W
wangxinxin08 已提交
1147 1148 1149 1150 1151
                if 'gt_keypoint' in sample.keys():
                    sample['gt_keypoint'] = gt_keypoints[0]
                    sample['keypoint_ignore'] = gt_keypoints[1]
                return sample
            return sample
Q
qingqing01 已提交
1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254


@register_op
class RandomCrop(BaseOperator):
    """Random crop image and bboxes.
    Args:
        aspect_ratio (list): aspect ratio of cropped region.
            in [min, max] format.
        thresholds (list): iou thresholds for decide a valid bbox crop.
        scaling (list): ratio between a cropped region and the original image.
             in [min, max] format.
        num_attempts (int): number of tries before giving up.
        allow_no_crop (bool): allow return without actually cropping them.
        cover_all_box (bool): ensure all bboxes are covered in the final crop.
        is_mask_crop(bool): whether crop the segmentation.
    """

    def __init__(self,
                 aspect_ratio=[.5, 2.],
                 thresholds=[.0, .1, .3, .5, .7, .9],
                 scaling=[.3, 1.],
                 num_attempts=50,
                 allow_no_crop=True,
                 cover_all_box=False,
                 is_mask_crop=False):
        super(RandomCrop, self).__init__()
        self.aspect_ratio = aspect_ratio
        self.thresholds = thresholds
        self.scaling = scaling
        self.num_attempts = num_attempts
        self.allow_no_crop = allow_no_crop
        self.cover_all_box = cover_all_box
        self.is_mask_crop = is_mask_crop

    def crop_segms(self, segms, valid_ids, crop, height, width):
        def _crop_poly(segm, crop):
            xmin, ymin, xmax, ymax = crop
            crop_coord = [xmin, ymin, xmin, ymax, xmax, ymax, xmax, ymin]
            crop_p = np.array(crop_coord).reshape(4, 2)
            crop_p = Polygon(crop_p)

            crop_segm = list()
            for poly in segm:
                poly = np.array(poly).reshape(len(poly) // 2, 2)
                polygon = Polygon(poly)
                if not polygon.is_valid:
                    exterior = polygon.exterior
                    multi_lines = exterior.intersection(exterior)
                    polygons = shapely.ops.polygonize(multi_lines)
                    polygon = MultiPolygon(polygons)
                multi_polygon = list()
                if isinstance(polygon, MultiPolygon):
                    multi_polygon = copy.deepcopy(polygon)
                else:
                    multi_polygon.append(copy.deepcopy(polygon))
                for per_polygon in multi_polygon:
                    inter = per_polygon.intersection(crop_p)
                    if not inter:
                        continue
                    if isinstance(inter, (MultiPolygon, GeometryCollection)):
                        for part in inter:
                            if not isinstance(part, Polygon):
                                continue
                            part = np.squeeze(
                                np.array(part.exterior.coords[:-1]).reshape(1,
                                                                            -1))
                            part[0::2] -= xmin
                            part[1::2] -= ymin
                            crop_segm.append(part.tolist())
                    elif isinstance(inter, Polygon):
                        crop_poly = np.squeeze(
                            np.array(inter.exterior.coords[:-1]).reshape(1, -1))
                        crop_poly[0::2] -= xmin
                        crop_poly[1::2] -= ymin
                        crop_segm.append(crop_poly.tolist())
                    else:
                        continue
            return crop_segm

        def _crop_rle(rle, crop, height, width):
            if 'counts' in rle and type(rle['counts']) == list:
                rle = mask_util.frPyObjects(rle, height, width)
            mask = mask_util.decode(rle)
            mask = mask[crop[1]:crop[3], crop[0]:crop[2]]
            rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
            return rle

        crop_segms = []
        for id in valid_ids:
            segm = segms[id]
            if is_poly(segm):
                import copy
                import shapely.ops
                from shapely.geometry import Polygon, MultiPolygon, GeometryCollection
                logging.getLogger("shapely").setLevel(logging.WARNING)
                # Polygon format
                crop_segms.append(_crop_poly(segm, crop))
            else:
                # RLE format
                import pycocotools.mask as mask_util
                crop_segms.append(_crop_rle(segm, crop, height, width))
        return crop_segms

W
wangxinxin08 已提交
1255
    def apply(self, sample, context=None):
Q
qingqing01 已提交
1256 1257 1258
        if 'gt_bbox' in sample and len(sample['gt_bbox']) == 0:
            return sample

W
wangxinxin08 已提交
1259
        h, w = sample['image'].shape[:2]
Q
qingqing01 已提交
1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341
        gt_bbox = sample['gt_bbox']

        # NOTE Original method attempts to generate one candidate for each
        # threshold then randomly sample one from the resulting list.
        # Here a short circuit approach is taken, i.e., randomly choose a
        # threshold and attempt to find a valid crop, and simply return the
        # first one found.
        # The probability is not exactly the same, kinda resembling the
        # "Monty Hall" problem. Actually carrying out the attempts will affect
        # observability (just like opening doors in the "Monty Hall" game).
        thresholds = list(self.thresholds)
        if self.allow_no_crop:
            thresholds.append('no_crop')
        np.random.shuffle(thresholds)

        for thresh in thresholds:
            if thresh == 'no_crop':
                return sample

            found = False
            for i in range(self.num_attempts):
                scale = np.random.uniform(*self.scaling)
                if self.aspect_ratio is not None:
                    min_ar, max_ar = self.aspect_ratio
                    aspect_ratio = np.random.uniform(
                        max(min_ar, scale**2), min(max_ar, scale**-2))
                    h_scale = scale / np.sqrt(aspect_ratio)
                    w_scale = scale * np.sqrt(aspect_ratio)
                else:
                    h_scale = np.random.uniform(*self.scaling)
                    w_scale = np.random.uniform(*self.scaling)
                crop_h = h * h_scale
                crop_w = w * w_scale
                if self.aspect_ratio is None:
                    if crop_h / crop_w < 0.5 or crop_h / crop_w > 2.0:
                        continue

                crop_h = int(crop_h)
                crop_w = int(crop_w)
                crop_y = np.random.randint(0, h - crop_h)
                crop_x = np.random.randint(0, w - crop_w)
                crop_box = [crop_x, crop_y, crop_x + crop_w, crop_y + crop_h]
                iou = self._iou_matrix(
                    gt_bbox, np.array(
                        [crop_box], dtype=np.float32))
                if iou.max() < thresh:
                    continue

                if self.cover_all_box and iou.min() < thresh:
                    continue

                cropped_box, valid_ids = self._crop_box_with_center_constraint(
                    gt_bbox, np.array(
                        crop_box, dtype=np.float32))
                if valid_ids.size > 0:
                    found = True
                    break

            if found:
                if self.is_mask_crop and 'gt_poly' in sample and len(sample[
                        'gt_poly']) > 0:
                    crop_polys = self.crop_segms(
                        sample['gt_poly'],
                        valid_ids,
                        np.array(
                            crop_box, dtype=np.int64),
                        h,
                        w)
                    if [] in crop_polys:
                        delete_id = list()
                        valid_polys = list()
                        for id, crop_poly in enumerate(crop_polys):
                            if crop_poly == []:
                                delete_id.append(id)
                            else:
                                valid_polys.append(crop_poly)
                        valid_ids = np.delete(valid_ids, delete_id)
                        if len(valid_polys) == 0:
                            return sample
                        sample['gt_poly'] = valid_polys
                    else:
                        sample['gt_poly'] = crop_polys
W
wangxinxin08 已提交
1342 1343 1344 1345 1346 1347 1348

                if 'gt_segm' in sample:
                    sample['gt_segm'] = self._crop_segm(sample['gt_segm'],
                                                        crop_box)
                    sample['gt_segm'] = np.take(
                        sample['gt_segm'], valid_ids, axis=0)

Q
qingqing01 已提交
1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393
                sample['image'] = self._crop_image(sample['image'], crop_box)
                sample['gt_bbox'] = np.take(cropped_box, valid_ids, axis=0)
                sample['gt_class'] = np.take(
                    sample['gt_class'], valid_ids, axis=0)
                if 'gt_score' in sample:
                    sample['gt_score'] = np.take(
                        sample['gt_score'], valid_ids, axis=0)

                if 'is_crowd' in sample:
                    sample['is_crowd'] = np.take(
                        sample['is_crowd'], valid_ids, axis=0)
                return sample

        return sample

    def _iou_matrix(self, a, b):
        tl_i = np.maximum(a[:, np.newaxis, :2], b[:, :2])
        br_i = np.minimum(a[:, np.newaxis, 2:], b[:, 2:])

        area_i = np.prod(br_i - tl_i, axis=2) * (tl_i < br_i).all(axis=2)
        area_a = np.prod(a[:, 2:] - a[:, :2], axis=1)
        area_b = np.prod(b[:, 2:] - b[:, :2], axis=1)
        area_o = (area_a[:, np.newaxis] + area_b - area_i)
        return area_i / (area_o + 1e-10)

    def _crop_box_with_center_constraint(self, box, crop):
        cropped_box = box.copy()

        cropped_box[:, :2] = np.maximum(box[:, :2], crop[:2])
        cropped_box[:, 2:] = np.minimum(box[:, 2:], crop[2:])
        cropped_box[:, :2] -= crop[:2]
        cropped_box[:, 2:] -= crop[:2]

        centers = (box[:, :2] + box[:, 2:]) / 2
        valid = np.logical_and(crop[:2] <= centers,
                               centers < crop[2:]).all(axis=1)
        valid = np.logical_and(
            valid, (cropped_box[:, :2] < cropped_box[:, 2:]).all(axis=1))

        return cropped_box, np.where(valid)[0]

    def _crop_image(self, img, crop):
        x1, y1, x2, y2 = crop
        return img[y1:y2, x1:x2, :]

W
wangxinxin08 已提交
1394 1395 1396
    def _crop_segm(self, segm, crop):
        x1, y1, x2, y2 = crop
        return segm[:, y1:y2, x1:x2]
Q
qingqing01 已提交
1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417


@register_op
class RandomScaledCrop(BaseOperator):
    """Resize image and bbox based on long side (with optional random scaling),
       then crop or pad image to target size.
    Args:
        target_dim (int): target size.
        scale_range (list): random scale range.
        interp (int): interpolation method, default to `cv2.INTER_LINEAR`.
    """

    def __init__(self,
                 target_dim=512,
                 scale_range=[.1, 2.],
                 interp=cv2.INTER_LINEAR):
        super(RandomScaledCrop, self).__init__()
        self.target_dim = target_dim
        self.scale_range = scale_range
        self.interp = interp

W
wangxinxin08 已提交
1418 1419 1420
    def apply(self, sample, context=None):
        img = sample['image']
        h, w = img.shape[:2]
Q
qingqing01 已提交
1421 1422 1423 1424 1425
        random_scale = np.random.uniform(*self.scale_range)
        dim = self.target_dim
        random_dim = int(dim * random_scale)
        dim_max = max(h, w)
        scale = random_dim / dim_max
W
wangxinxin08 已提交
1426 1427
        resize_w = w * scale
        resize_h = h * scale
Q
qingqing01 已提交
1428 1429
        offset_x = int(max(0, np.random.uniform(0., resize_w - dim)))
        offset_y = int(max(0, np.random.uniform(0., resize_h - dim)))
W
wangxinxin08 已提交
1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442

        img = cv2.resize(img, (resize_w, resize_h), interpolation=self.interp)
        img = np.array(img)
        canvas = np.zeros((dim, dim, 3), dtype=img.dtype)
        canvas[:min(dim, resize_h), :min(dim, resize_w), :] = img[
            offset_y:offset_y + dim, offset_x:offset_x + dim, :]
        sample['image'] = canvas
        sample['im_shape'] = np.asarray([resize_h, resize_w], dtype=np.float32)
        scale_factor = sample['sacle_factor']
        sample['scale_factor'] = np.asarray(
            [scale_factor[0] * scale, scale_factor[1] * scale],
            dtype=np.float32)

Q
qingqing01 已提交
1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457
        if 'gt_bbox' in sample and len(sample['gt_bbox']) > 0:
            scale_array = np.array([scale, scale] * 2, dtype=np.float32)
            shift_array = np.array([offset_x, offset_y] * 2, dtype=np.float32)
            boxes = sample['gt_bbox'] * scale_array - shift_array
            boxes = np.clip(boxes, 0, dim - 1)
            # filter boxes with no area
            area = np.prod(boxes[..., 2:] - boxes[..., :2], axis=1)
            valid = (area > 1.).nonzero()[0]
            sample['gt_bbox'] = boxes[valid]
            sample['gt_class'] = sample['gt_class'][valid]

        return sample


@register_op
W
wangxinxin08 已提交
1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473
class Cutmix(BaseOperator):
    def __init__(self, alpha=1.5, beta=1.5):
        """ 
        CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features, see https://arxiv.org/abs/1905.04899
        Cutmix image and gt_bbbox/gt_score
        Args:
             alpha (float): alpha parameter of beta distribute
             beta (float): beta parameter of beta distribute
        """
        super(Cutmix, self).__init__()
        self.alpha = alpha
        self.beta = beta
        if self.alpha <= 0.0:
            raise ValueError("alpha shold be positive in {}".format(self))
        if self.beta <= 0.0:
            raise ValueError("beta shold be positive in {}".format(self))
Q
qingqing01 已提交
1474

W
wangxinxin08 已提交
1475 1476 1477 1478 1479
    def apply_image(self, img1, img2, factor):
        """ _rand_bbox """
        h = max(img1.shape[0], img2.shape[0])
        w = max(img1.shape[1], img2.shape[1])
        cut_rat = np.sqrt(1. - factor)
Q
qingqing01 已提交
1480

W
wangxinxin08 已提交
1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531
        cut_w = np.int(w * cut_rat)
        cut_h = np.int(h * cut_rat)

        # uniform
        cx = np.random.randint(w)
        cy = np.random.randint(h)

        bbx1 = np.clip(cx - cut_w // 2, 0, w - 1)
        bby1 = np.clip(cy - cut_h // 2, 0, h - 1)
        bbx2 = np.clip(cx + cut_w // 2, 0, w - 1)
        bby2 = np.clip(cy + cut_h // 2, 0, h - 1)

        img_1 = np.zeros((h, w, img1.shape[2]), 'float32')
        img_1[:img1.shape[0], :img1.shape[1], :] = \
            img1.astype('float32')
        img_2 = np.zeros((h, w, img2.shape[2]), 'float32')
        img_2[:img2.shape[0], :img2.shape[1], :] = \
            img2.astype('float32')
        img_1[bby1:bby2, bbx1:bbx2, :] = img2[bby1:bby2, bbx1:bbx2, :]
        return img_1

    def __call__(self, sample, context=None):
        if not isinstance(sample, Sequence):
            return sample

        assert len(sample) == 2, 'cutmix need two samples'

        factor = np.random.beta(self.alpha, self.beta)
        factor = max(0.0, min(1.0, factor))
        if factor >= 1.0:
            return sample[0]
        if factor <= 0.0:
            return sample[1]
        img1 = sample[0]['image']
        img2 = sample[1]['image']
        img = self.apply_image(img1, img2, factor)
        gt_bbox1 = sample[0]['gt_bbox']
        gt_bbox2 = sample[1]['gt_bbox']
        gt_bbox = np.concatenate((gt_bbox1, gt_bbox2), axis=0)
        gt_class1 = sample[0]['gt_class']
        gt_class2 = sample[1]['gt_class']
        gt_class = np.concatenate((gt_class1, gt_class2), axis=0)
        gt_score1 = sample[0]['gt_score']
        gt_score2 = sample[1]['gt_score']
        gt_score = np.concatenate(
            (gt_score1 * factor, gt_score2 * (1. - factor)), axis=0)
        sample = sample[0]
        sample['image'] = img
        sample['gt_bbox'] = gt_bbox
        sample['gt_score'] = gt_score
        sample['gt_class'] = gt_class
Q
qingqing01 已提交
1532 1533 1534 1535
        return sample


@register_op
W
wangxinxin08 已提交
1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549
class Mixup(BaseOperator):
    def __init__(self, alpha=1.5, beta=1.5):
        """ Mixup image and gt_bbbox/gt_score
        Args:
            alpha (float): alpha parameter of beta distribute
            beta (float): beta parameter of beta distribute
        """
        super(Mixup, self).__init__()
        self.alpha = alpha
        self.beta = beta
        if self.alpha <= 0.0:
            raise ValueError("alpha shold be positive in {}".format(self))
        if self.beta <= 0.0:
            raise ValueError("beta shold be positive in {}".format(self))
Q
qingqing01 已提交
1550

W
wangxinxin08 已提交
1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562
    def apply_image(self, img1, img2, factor):
        h = max(img1.shape[0], img2.shape[0])
        w = max(img1.shape[1], img2.shape[1])
        img = np.zeros((h, w, img1.shape[2]), 'float32')
        img[:img1.shape[0], :img1.shape[1], :] = \
            img1.astype('float32') * factor
        img[:img2.shape[0], :img2.shape[1], :] += \
            img2.astype('float32') * (1.0 - factor)
        return img.astype('uint8')

    def __call__(self, sample, context=None):
        if not isinstance(sample, Sequence):
Q
qingqing01 已提交
1563 1564
            return sample

W
wangxinxin08 已提交
1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666
        assert len(sample) == 2, 'mixup need two samples'

        factor = np.random.beta(self.alpha, self.beta)
        factor = max(0.0, min(1.0, factor))
        if factor >= 1.0:
            return sample[0]
        if factor <= 0.0:
            return sample[1]
        im = self.apply_image(sample[0]['image'], sample[1]['image'], factor)
        result = copy.deepcopy(sample[0])
        result['image'] = im
        # apply bbox and score
        if 'gt_bbox' in sample[0]:
            gt_bbox1 = sample[0]['gt_bbox']
            gt_bbox2 = sample[1]['gt_bbox']
            gt_bbox = np.concatenate((gt_bbox1, gt_bbox2), axis=0)
            result['gt_bbox'] = gt_bbox
        if 'gt_class' in sample[0]:
            gt_class1 = sample[0]['gt_class']
            gt_class2 = sample[1]['gt_class']
            gt_class = np.concatenate((gt_class1, gt_class2), axis=0)
            result['gt_class'] = gt_class

            gt_score1 = np.ones_like(sample[0]['gt_class'])
            gt_score2 = np.ones_like(sample[1]['gt_class'])
            gt_score = np.concatenate(
                (gt_score1 * factor, gt_score2 * (1. - factor)), axis=0)
            result['gt_score'] = gt_score
        if 'is_crowd' in sample[0]:
            is_crowd1 = sample[0]['is_crowd']
            is_crowd2 = sample[1]['is_crowd']
            is_crowd = np.concatenate((is_crowd1, is_crowd2), axis=0)
            result['is_crowd'] = is_crowd
        if 'difficult' in sample[0]:
            is_difficult1 = sample[0]['difficult']
            is_difficult2 = sample[1]['difficult']
            is_difficult = np.concatenate(
                (is_difficult1, is_difficult2), axis=0)
            result['difficult'] = is_difficult

        return result


@register_op
class NormalizeBox(BaseOperator):
    """Transform the bounding box's coornidates to [0,1]."""

    def __init__(self):
        super(NormalizeBox, self).__init__()

    def apply(self, sample, context):
        im = sample['image']
        gt_bbox = sample['gt_bbox']
        height, width, _ = im.shape
        for i in range(gt_bbox.shape[0]):
            gt_bbox[i][0] = gt_bbox[i][0] / width
            gt_bbox[i][1] = gt_bbox[i][1] / height
            gt_bbox[i][2] = gt_bbox[i][2] / width
            gt_bbox[i][3] = gt_bbox[i][3] / height
        sample['gt_bbox'] = gt_bbox

        if 'gt_keypoint' in sample.keys():
            gt_keypoint = sample['gt_keypoint']

            for i in range(gt_keypoint.shape[1]):
                if i % 2:
                    gt_keypoint[:, i] = gt_keypoint[:, i] / height
                else:
                    gt_keypoint[:, i] = gt_keypoint[:, i] / width
            sample['gt_keypoint'] = gt_keypoint

        return sample


@register_op
class BboxXYXY2XYWH(BaseOperator):
    """
    Convert bbox XYXY format to XYWH format.
    """

    def __init__(self):
        super(BboxXYXY2XYWH, self).__init__()

    def apply(self, sample, context=None):
        assert 'gt_bbox' in sample
        bbox = sample['gt_bbox']
        bbox[:, 2:4] = bbox[:, 2:4] - bbox[:, :2]
        bbox[:, :2] = bbox[:, :2] + bbox[:, 2:4] / 2.
        sample['gt_bbox'] = bbox
        return sample


@register_op
class PadBox(BaseOperator):
    def __init__(self, num_max_boxes=50):
        """
        Pad zeros to bboxes if number of bboxes is less than num_max_boxes.
        Args:
            num_max_boxes (int): the max number of bboxes
        """
        self.num_max_boxes = num_max_boxes
        super(PadBox, self).__init__()
Q
qingqing01 已提交
1667

W
wangxinxin08 已提交
1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700
    def apply(self, sample, context=None):
        assert 'gt_bbox' in sample
        bbox = sample['gt_bbox']
        gt_num = min(self.num_max_boxes, len(bbox))
        num_max = self.num_max_boxes
        # fields = context['fields'] if context else []
        pad_bbox = np.zeros((num_max, 4), dtype=np.float32)
        if gt_num > 0:
            pad_bbox[:gt_num, :] = bbox[:gt_num, :]
        sample['gt_bbox'] = pad_bbox
        if 'gt_class' in sample:
            pad_class = np.zeros((num_max, ), dtype=np.int32)
            if gt_num > 0:
                pad_class[:gt_num] = sample['gt_class'][:gt_num, 0]
            sample['gt_class'] = pad_class
        if 'gt_score' in sample:
            pad_score = np.zeros((num_max, ), dtype=np.float32)
            if gt_num > 0:
                pad_score[:gt_num] = sample['gt_score'][:gt_num, 0]
            sample['gt_score'] = pad_score
        # in training, for example in op ExpandImage,
        # the bbox and gt_class is expandded, but the difficult is not,
        # so, judging by it's length
        if 'difficult' in sample:
            pad_diff = np.zeros((num_max, ), dtype=np.int32)
            if gt_num > 0:
                pad_diff[:gt_num] = sample['difficult'][:gt_num, 0]
            sample['difficult'] = pad_diff
        if 'is_crowd' in sample:
            pad_crowd = np.zeros((num_max, ), dtype=np.int32)
            if gt_num > 0:
                pad_crowd[:gt_num] = sample['is_crowd'][:gt_num, 0]
            sample['is_crowd'] = pad_crowd
Q
qingqing01 已提交
1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719
        return sample


@register_op
class DebugVisibleImage(BaseOperator):
    """
    In debug mode, visualize images according to `gt_box`.
    (Currently only supported when not cropping and flipping image.)
    """

    def __init__(self, output_dir='output/debug', is_normalized=False):
        super(DebugVisibleImage, self).__init__()
        self.is_normalized = is_normalized
        self.output_dir = output_dir
        if not os.path.isdir(output_dir):
            os.makedirs(output_dir)
        if not isinstance(self.is_normalized, bool):
            raise TypeError("{}: input type is invalid.".format(self))

W
wangxinxin08 已提交
1720
    def apply(self, sample, context=None):
Q
qingqing01 已提交
1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761
        image = Image.open(sample['im_file']).convert('RGB')
        out_file_name = sample['im_file'].split('/')[-1]
        width = sample['w']
        height = sample['h']
        gt_bbox = sample['gt_bbox']
        gt_class = sample['gt_class']
        draw = ImageDraw.Draw(image)
        for i in range(gt_bbox.shape[0]):
            if self.is_normalized:
                gt_bbox[i][0] = gt_bbox[i][0] * width
                gt_bbox[i][1] = gt_bbox[i][1] * height
                gt_bbox[i][2] = gt_bbox[i][2] * width
                gt_bbox[i][3] = gt_bbox[i][3] * height

            xmin, ymin, xmax, ymax = gt_bbox[i]
            draw.line(
                [(xmin, ymin), (xmin, ymax), (xmax, ymax), (xmax, ymin),
                 (xmin, ymin)],
                width=2,
                fill='green')
            # draw label
            text = str(gt_class[i][0])
            tw, th = draw.textsize(text)
            draw.rectangle(
                [(xmin + 1, ymin - th), (xmin + tw + 1, ymin)], fill='green')
            draw.text((xmin + 1, ymin - th), text, fill=(255, 255, 255))

        if 'gt_keypoint' in sample.keys():
            gt_keypoint = sample['gt_keypoint']
            if self.is_normalized:
                for i in range(gt_keypoint.shape[1]):
                    if i % 2:
                        gt_keypoint[:, i] = gt_keypoint[:, i] * height
                    else:
                        gt_keypoint[:, i] = gt_keypoint[:, i] * width
            for i in range(gt_keypoint.shape[0]):
                keypoint = gt_keypoint[i]
                for j in range(int(keypoint.shape[0] / 2)):
                    x1 = round(keypoint[2 * j]).astype(np.int32)
                    y1 = round(keypoint[2 * j + 1]).astype(np.int32)
                    draw.ellipse(
W
wangxinxin08 已提交
1762
                        (x1, y1, x1 + 5, y1 + 5), fill='green', outline='green')
Q
qingqing01 已提交
1763 1764 1765
        save_path = os.path.join(self.output_dir, out_file_name)
        image.save(save_path, quality=95)
        return sample
W
wangxinxin08 已提交
1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935


@register_op
class Pad(BaseOperator):
    def __init__(self,
                 size=None,
                 size_divisor=32,
                 pad_mode=0,
                 offsets=None,
                 fill_value=(127.5, 127.5, 127.5)):
        """
        Pad image to a specified size or multiple of size_divisor. random target_size and interpolation method
        Args:
            size (int, Sequence): image target size, if None, pad to multiple of size_divisor, default None
            size_divisor (int): size divisor, default 32
            pad_mode (int): pad mode, currently only supports four modes [-1, 0, 1, 2]. if -1, use specified offsets
                if 0, only pad to right and bottom. if 1, pad according to center. if 2, only pad left and top
            fill_value (bool): rgb value of pad area, default (127.5, 127.5, 127.5)
        """
        super(Pad, self).__init__()

        if not isinstance(size, (int, Sequence)):
            raise TypeError(
                "Type of target_size is invalid when random_size is True. \
                            Must be List, now is {}".format(type(size)))

        if isinstance(size, int):
            size = [size, size]

        assert pad_mode in [
            -1, 0, 1, 2
        ], 'currently only supports four modes [-1, 0, 1, 2]'
        assert pad_mode == -1 and offsets, 'if pad_mode is -1, offsets should not be None'

        self.size = size
        self.size_divisor = size_divisor
        self.pad_mode = pad_mode
        self.fill_value = fill_value
        self.offsets = offsets

    def apply_segm(self, segms, offsets, im_size, size):
        def _expand_poly(poly, x, y):
            expanded_poly = np.array(poly)
            expanded_poly[0::2] += x
            expanded_poly[1::2] += y
            return expanded_poly.tolist()

        def _expand_rle(rle, x, y, height, width, h, w):
            if 'counts' in rle and type(rle['counts']) == list:
                rle = mask_util.frPyObjects(rle, height, width)
            mask = mask_util.decode(rle)
            expanded_mask = np.full((h, w), 0).astype(mask.dtype)
            expanded_mask[y:y + height, x:x + width] = mask
            rle = mask_util.encode(
                np.array(
                    expanded_mask, order='F', dtype=np.uint8))
            return rle

        x, y = offsets
        height, width = im_size
        h, w = size
        expanded_segms = []
        for segm in segms:
            if is_poly(segm):
                # Polygon format
                expanded_segms.append(
                    [_expand_poly(poly, x, y) for poly in segm])
            else:
                # RLE format
                import pycocotools.mask as mask_util
                expanded_segms.append(
                    _expand_rle(segm, x, y, height, width, h, w))
        return expanded_segms

    def apply_bbox(self, bbox, offsets):
        return bbox + np.array(offsets * 2, dtype=np.float32)

    def apply_keypoint(self, keypoints, offsets):
        n = len(keypoints[0]) // 2
        return keypoints + np.array(offsets * n, dtype=np.float32)

    def apply_image(self, image, offsets, im_size, size):
        x, y = offsets
        im_h, im_w = im_size
        h, w = size
        canvas = np.ones((h, w, 3), dtype=np.float32)
        canvas *= np.array(self.fill_value, dtype=np.float32)
        canvas[y:y + im_h, x:x + im_w, :] = image.astype(np.float32)
        return canvas

    def apply(self, sample, context=None):
        im = sample['image']
        im_h, im_w = im.shape[:2]
        if self.size:
            h, w = self.size
            assert (
                im_h < h and im_w < w
            ), '(h, w) of target size should be greater than (im_h, im_w)'
        else:
            h = np.ceil(im_h // self.size_divisor) * self.size_divisor
            w = np.ceil(im_w / self.size_divisor) * self.size_divisor

        if h == im_h and w == im_w:
            return sample

        if self.pad_mode == -1:
            offset_x, offset_y = self.offsets
        elif self.pad_mode == 0:
            offset_y, offset_x = 0, 0
        elif self.pad_mode == 1:
            offset_y, offset_x = (h - im_h) // 2, (w - im_w) // 2
        else:
            offset_y, offset_x = h - im_h, w - im_w

        offsets, im_size, size = [offset_x, offset_y], [im_h, im_w], [h, w]

        sample['image'] = self.apply_image(im, offsets, im_size, size)

        if self.pad_mode == 0:
            return sample
        if 'gt_bbox' in sample and len(sample['gt_bbox']) > 0:
            sample['gt_bbox'] = self.apply_bbox(sample['gt_bbox'], offsets)

        if 'gt_poly' in sample and len(sample['gt_poly']) > 0:
            sample['gt_poly'] = self.apply_segm(sample['gt_poly'], offsets,
                                                im_size, size)

        if 'gt_keypoint' in sample and len(sample['gt_keypoint']) > 0:
            sample['gt_keypoint'] = self.apply_keypoint(sample['gt_keypoint'],
                                                        offsets)

        return sample


@register_op
class Poly2Mask(BaseOperator):
    """
    gt poly to mask annotations
    """

    def __init__(self):
        super(Poly2Mask, self).__init__()
        import pycocotools.mask as maskUtils
        self.maskutils = maskUtils

    def _poly2mask(self, mask_ann, img_h, img_w):
        if isinstance(mask_ann, list):
            # polygon -- a single object might consist of multiple parts
            # we merge all parts into one mask rle code
            rles = self.maskutils.frPyObjects(mask_ann, img_h, img_w)
            rle = self.maskutils.merge(rles)
        elif isinstance(mask_ann['counts'], list):
            # uncompressed RLE
            rle = self.maskutils.frPyObjects(mask_ann, img_h, img_w)
        else:
            # rle
            rle = mask_ann
        mask = self.maskutils.decode(rle)
        return mask

    def apply(self, sample, context=None):
        assert 'gt_poly' in sample
        im_h = sample['h']
        im_w = sample['w']
        masks = [
            self._poly2mask(gt_poly, im_h, im_w)
            for gt_poly in sample['gt_poly']
        ]
        sample['gt_segm'] = np.asarray(masks).astype(np.uint8)
        return sample